Exploiting DOLCE, SUMO-OWL, and OpenCyc to Boost the Ontology Matching Process Technical Report DISI-TR-08-08, Revised Version, Computer Science Department, DISI, Università degli Studi di Genova, Italy, 2008 Viviana Mascardi DISI, Università degli Studi di Genova, Via Dodecaneso 35, 16146, Genova, Italy, E-mail:
[email protected] Angela Locoro DIBE, Università degli Studi di Genova, Via All’Opera Pia 11a, 16145 Genova, Italy, E-mail:
[email protected] Paolo Rosso Natural Language Engineering Lab., DSIC, Universidad Politécnica de Valencia, Camino de Vera s/n, 46022, Valencia Spain, E-mail:
[email protected] Abstract This technical report describes a set of algorithms for exploiting upper ontologies (DOLCE, SUMO-OWL, and OpenCyc) as bridges in the ontology matching process, and discusses experimental results. The experiments demonstrate that when our “structural matching method via upper ontology” uses an upper ontology with a sufficient number of entities (OpenCyc, SUMO-OWL), the recall is significantly improved and the precision is kept (at least with OpenCyc). Instead, the “non structural matching method” via OpenCyc and via SUMO-OWL gives a precision greater than the one obtained with the “non structural direct alignment”, and a comparable recall. The “mixed method” that combines the results of both structural direct alignment and structural alignment via upper ontologies, gives the best average recall and F-measure.
1
1
Introduction
“Ontology matching” is the process of finding correspondences between entities belonging to two or more ontologies. The output of this process, named “ontology alignment”, may be computed by using different techniques and is used to cope with the semantic and syntactic heterogeneity of information represented by means of ontologies, in order to identify couples of entities with related meaning, belonging to two different ontologies. An upper ontology is “an attempt to create an ontology which describes very general concepts that are the same across all domains” [15]. Few upper ontologies exist: BFO [7], Cyc [9], DOLCE [6], GFO [8], PROTON [3], Sowa’s ontology [12], and SUMO [10]. They vary in dimension, ranging from the 30 classes of Sowa’s ontology to the 300,000 of Cyc, in representation language (OWL [14], KIF [1], First Order Logic), in structure (monolithic vs. decomposed into modules), and in developed applications. Nevertheless, all of them describe general concepts (also named “classes”) and share the aim to have a large number of ontologies accessible under them. Some of them, such as Cyc and SUMO, also provide domain-dependent sub-ontologies integrated with the upper one. This technical report describes a set of algorithms for exploiting upper ontologies as bridges in the ontology matching process and discusses experimental results. We have chosen to run our experiments with DOLCE, SUMO-OWL (a restricted version of SUMO translated into OWL), and OpenCyc (the open version of Cyc), because they are the largest and most used upper ontologies. The experiments demonstrate that when our “structural matching method via upper ontology” uses an upper ontology with a sufficient number of entities (OpenCyc, SUMO-OWL), the recall is significantly improved and the precision is kept (at least with OpenCyc). Instead, the “non structural matching method” via OpenCyc and via SUMO-OWL gives a precision greater than the one obtained with the “non structural direct alignment”, and a comparable recall. The “mixed method” that combines the results of both structural direct alignment and structural alignment via upper ontologies, gives the best average recall and F-measure. The report is organised in the following way: Section 2 describes the algorithms we have designed for implementing the ontology matching via upper ontologies, Section 3 describes the tests we have carried out and their results, and Section 4 concludes.
2
Our Algorithms for Ontology Matching via Upper Ontologies
In order to use upper ontologies as bridges in the ontology matching process, we implemented many algorithms. U o_match is based on three functions: aggregate(a, a0 ), parallel_match(o, o0 , res, par), and merge(a, a0 ), where o and o0 are ontologies, a and a0 are alignments, res are external resources, and par are parameters. All the three functions return an alignment. Aggregate(a, a0 ) produces the alignment obtained by making the union of all the correspondences in a and a0 , and choosing the correspondence with highest confidence 2
measure, in case the same correspondence1 belongs to both a and a0 . P arallel_match(o, o0 , res, par) computes an alignment between o and o0 by applying substring, n-gram [2], SMOA [13], and language-based methods in parallel, as suggested in [5, Chap. 5.1], and aggregating them. The only external resource we use is WordNet (res = {W ordN et}), which is given in input to the language-based method, and the only parameter is a configurable threshold in [0, 1] used for discarding correspondences that are not relevant, since their confidence is lower than it (par = {th}). In order to simulate the parallelism of the aggregation process, aggregate is initially called on the first two alignments obtained by running the first two matching algorithms; the output is then aggregated with the third alignment, and so on. M erge(a, a0 ) computes the alignment a00 in such a way that a correspondence hid, c, c0 , r, conf i belongs to a00 iff ∃ cu such that hid1 , c, cu , r, conf1 i ∈ a, hid2 , c0 , cu , r, conf2 i ∈ a0 , and conf = conf1 ∗ conf2 . In our algorithm the inputs of merge, a and a0 , are alignments between o and the upper ontology u, and o0 and u, respectively. Thus, the second concept cu in the correspondences belongs to u. If both c ∈ o and c0 ∈ o0 correspond to the same concept cu ∈ u, then c and c0 are related via r with confidence conf1 ∗ conf2 . The choice of conf1 ∗ conf2 as confidence of the resulting alignment ensures that the confidence remains in [0, 1], and that initial high confidences lead to a resulting confidence which is still high, whereas at least one low confidence leads to a low resulting confidence. The uo_match(o, o0 , u, res, par) function just calls merge (parallel_match(o, u, res, par), parallel_match(o0 , u, res, par)) with res = {W ordN et} and par = {th}. The th parameter has been set to 0.5 in our experiments, and u is an upper ontology. Besides the parallel_match function, we have implemented the structural_parallel _match function that adds to the output alignment those correspondences hid, c, c0 , r, conf i such that • either c is identical to one of c0 ’s super-concepts (or vice versa, c0 is identical to one of the super-concepts of c), • or c and c0 have two super-concepts, say s ∈ o and s0 ∈ o0 respectively, identical. In a similar way, we have implemented a structural_merge function that merges two alignments considering correspondences involving super-concepts into account. We have prefixed the names of these functions with structural because they look at the structure of the matched ontologies. Structural_merge takes a decay factor df ∈ [0, 1], used to measure the “confidence decay” as we consider relations that involve super-concepts. In our experiments, df has been set to 0.5. It also takes the upper ontology u used as the reference ontology for computing a and a0 as input, since it must be navigated in order to find the super-concepts of a given concept. 1 The
“same” apart from the correspondence identifier, which usually will not be the same.
3
Structural_merge(a, a0 , u, df ) computes the alignment a00 between o and o0 via u in such a way that a correspondence hid, c, c0 , r, conf i belongs to a00 if either • ∃ cu such that hid1 , c, cu , r, conf1 i ∈ a, hid2 , c0 , cu , r, conf2 i ∈ a0 , and conf = conf1 ∗ conf2 , or • ∃ cu , c0u such that hid1 , c, cu , r, conf1 i ∈ a, hid2 , c0 , c0u , r, conf2 i ∈ a0 , c0u is a super-concept of cu in u, and conf = conf1 ∗ conf2 ∗ df (note the confidence decay multiplicative factor), and vice versa. • ∃ cu , c0u such that hid1 , c, cu , r, conf1 i ∈ a, hid2 , c0 , c0u , r, conf2 i ∈ a0 , cu and c0u have a common super-concept in u, and conf = conf1 ∗ conf2 ∗ df 2 (note the power factor applied to the confidence decay). The structural_uo_match function is defined as structural_merge( parallel_match(o, u, W ordN et, th), parallel_match(o0 , u, W ordN et, th), {u}, {df }). The structural_uo_match uses the parallel_match function for computing the alignments, and not the structural_parallel_match one. The exploitation of the structure is demanded to the merging stage. The reason why we have also defined a structural_parallel_match function is that we want to compare the matching via upper ontology results and the direct match results in the case where both functions exploit structural information, and in the case where both do not. Thus, in our experiments, we have compared the results of uo_match with those of parallel_match, and the results of structural_uo_match with those of structural_parallel_match. For gaining in efficiency by reusing the results obtained by running uo_match, structural_uo_match has been implemented in the following way: • it creates an alignment between o and o0 via u by looking for mappings between – concepts belonging to o and super-concepts of concepts belonging o0 ; – vice versa, concepts belonging to o0 and super-concepts of concepts belonging o; – super-concepts of concepts in o and super-concepts of concepts in o0 . • it aggregates the alignment obtained with the uo_match, and the alignment obtained as above. A mixed_match algorithm obtained by aggregating the alignment output by the structural_parallel_match algorithm (direct matching exploiting structure), and the one output by the structural_uo_match algorithm (matching via upper ontology, exploiting structure), has also been implemented. Our algorithms accept input ontologies expressed in OWL. For their implementation we have extended the Align API version 3.1, delivered on February, the 5th, 2008, 4
and available form http://alignapi.gforge.inria.fr/ under GNU Lesser General Public License. Among the methods offered by this API, we have used: • StringDistAlignment, that provides the subStringDistance, ngramDistance, and smoaDistance string metric methods. • JWNLAlignment, that computes a substring distance between the entity names of the first ontology and the entity names of the second ontology expanded with WordNet 2.0 synsets. We have implemented a SubSupClassAlignment method that finds correspondences between c ∈ o and c0 ∈ o0 by looking at string equality between one of them, and one super-concept of the other one, or between super-concepts of both. We have also implemented the methods for alignment aggregation and merge. We created reference alignments that only match concepts, and we discarded correspondences between individuals and between properties from the alignments computed by our algorithms. The reason of our choice is that finding correspondences between properties in an accurate way would require to take their domain and range into account. The Align API version 3.1 that we used for computing similarity measures cannot take domain, range, and class into account in the correct way. This also explains our choice of discarding property restrictions when exploring the sub- and super-classes, and considering only subClassOf relations with simple entities used as their property values. In order to discard property restrictions when exploring the sub- and super-classes, we pre-processed the ontologies given in input to our algorithms, and eliminated property restrictions. Also, the easiest and most efficient way to discard correspondences involving properties and individuals from the alignments computed by our algorithms, was to pre-process the input ontologies and eliminate properties and individuals from them. We used the JENA OWL parser, http://jena.sourceforge.net/, for the pre-processing activities.
3
Experimentation
As indicators for measuring how good an alignment is, we have used precision, recall and F-measure adapted for ontology alignment evaluation [4]. The methodology that we have followed for carrying out our tests is aimed at ensuring the reproducibility of our experiments by reusing existing ontologies and existing APIs. It may be summarised in the following steps: 1. By exploiting the SWOOGLE Semantic Web Search Engine, http://swoogle. umbc.edu/, we have chosen 17 ontologies represented in OWL and available online. Only one of them has been reduced by hand to make it more tractable. The other 16 ones are exactly those that can be downloaded from the URLs listed in Table 1, and that we last accessed on February, 25th, 2008. 2. We have chosen 3 upper ontologies to use in our tests: SUMO-OWL, Cyc (in its open version, OpenCyc) and DOLCE. We have downloaded the OWL versions 5
of OpenCyc and DOLCE from http://www.cyc.com/2004/06/04/cyc and http://www.loa-cnr.it/ontologies/DLP_397.owl respectively. Our last access dates back to February, 25th, 2008. We have used an OWL translation of the SUO-KIF implementation of SUMO available from http://www. ontologyportal.org. The translation was performed using Sigma [11]. It contains restricted versions of SUMO, MILO, and all the domain ontologies but the terrorism and airport ones. SUMO-OWL is an order of magnitude smaller than the original SUMO represented in KIF, if one just counts terms, and it is even smaller if one looks at all the axioms necessarily lost in any translation from KIF to OWL. Our experiments do not involve the original SUMO ontology, but a lossy translation of it. 3. Using the JENA OWL parser, we have pre-processed all the 17 chosen ontologies and the 3 upper ontologies, and we have eliminated property restrictions, individuals, and properties. 4. We have designed the 10 tests to run, each one consisting of two ontologies to match and we have created the reference alignment for each test. All the reference alignments are available at http://www.disi.unige.it/person/ MascardiV/Software/OntologyMatchingViaUpperOntology.html. 5. For each test, we have run the following algorithms: • direct alignment without exploiting structure; • direct alignment exploiting structure; • alignment via SUMO-OWL, OpenCyc, and DOLCE without exploiting structure; • alignment via SUMO-OWL, OpenCyc, and DOLCE exploiting structure; • mixed alignment obtained by aggregating the structural alignment via SUMOOWL, OpenCyc, and DOLCE and the structural direct one. 6. For each test, and for each algorithm run within the test, we have computed: • number of correspondences found by the algorithm; • number of correct correspondences found by the algorithm; • precision; • recall; • F-measure. We exploited the library of evaluators provided by the Align API 3.1 for computing them. 7. We have aggregated the obtained results by identifying, for each test, the algorithms that allow us to get the best precision, the best recall, and the best F-measure, and by computing the average advantage of using upper ontologies.
6
Table 1 provides a summary of the ontologies used in our tests, including the upper ones. C stands for the total number of concepts, O stands for object properties, namely properties that link individuals to individuals, D stands for data properties, namely properties that link individuals to data values, and I stands for the number of individuals, or instances of the classes. Table 2 describes the algorithms that allowed us to obtain the best precision, the best recall, and the best F-measure for each test. Dir stands for “direct alignment”, NS means “no structure”, S means “with structure”, and M means “mixed”. Sometimes, different algorithms led to the same result. In those cases we listed them all. In seven tests out of ten, the best precision was obtained by methods that exploit non structural matching via upper ontologies. The best recall is always obtained by the mixed method. In two tests, the same best recall was obtained by the exploitation of structural matching via upper ontologies. The best F-measure is always obtained by methods that exploit upper ontologies. In only one test, the same best F-measure was obtained by direct matching methods too. The exploitation of OpenCyc ensures a very good precision, while good recall and F-measures can be obtained by both OpenCyc and SUMO-OWL in a comparable number of tests. Table 3 synthesises the average advantage in using upper ontologies vs. not using them. Each column refers to the comparison of the results obtained by using a given upper ontology (first element of the column’s name) and a given method (second element of the column’s name: again, NS means “no structure”, S means “with structure”, and M means “mixed”), and those obtained by performing the direct alignment with the same method (with or without structure respectively). For example, cell ((DOLCE, NS), Precision) reports the average difference in precision between aligning ontologies using DOLCE without exploiting the ontology structure, and performing the direct alignment without exploiting the structure. If we identify the precision obtained using DOLCE without structure in experiment i with p(DOLCE, N oStruct, i), and the precision obtained by performing the direct alignment without structure in experiment i with p(Direct, N oStruct, i), then
(DOLCE, NS, Prec.) =
Σ10 i=1 (p(DOLCE, N oStruct, i) − p(Direct, N oStruct, i)) 10
When the “mixed method” is used, we have compared its results with those of the direct alignment that obtains the best F-measure on that test.
4
Conclusions
In this paper we described a set of algorithms for exploiting upper ontologies as bridges in the ontology matching process, and we discussed the results of our experiments. These experiments have been carried out following a rigorous methodology that makes them easily reproducible: the three upper ontologies that we chose, SUMO-OWL,
7
OpenCyc and DOLCE, are available on the web, as well as the seventeen ontologies used as inputs for our ten tests, and the ten reference alignment we built by hand; our alignment algorithms extend an existing free API; all the assumptions and simplifications we made are clearly stated and motivated. The results of our tests may be summarised in the following way: 1. Very small upper ontologies like DOLCE do not possess enough information to properly act as bridges; their usage leads to very poor results. 2. OpenCyc and SUMO-OWL are large and detailed enough and give comparable results. OpenCyc performs better than SUMO-OWL, but algorithms that use it are less efficient than those that use SUMO-OWL, due to its dimensions. 3. If the usage scenario prevents the developer from performing a direct alignment, exploiting structural matching via upper ontologies is a valid alternative for improving the recall (with both SUMO-OWL and OpenCyc), keeping the same precision (with OpenCyc). Instead, if precision is more important than recall, the best matching method is the non structural one via OpenCyc and SUMO-OWL. 4. If the usage scenario allows the application developer to exploit both direct alignment and alignment via upper ontologies, the mixed method always gives the best recall, even if with a non negligible loss in precision. However, if we agree that F-measure synthesises the good features of the algorithm, mixed methods via SUMO-OWL and OpenCyc have the highest average F-measure, and thus should be considered the best ones. It must be noted that our results do not apply to the original SUMO upper ontology implemented in KIF, but to the SUMO-OWL version obtained by lossy translation. We could not compare the original SUMO directly with DOLCE and OpenCyc because we needed all input ontologies represented in the same language for running our experiments. Since SUMO in KIF is definitely larger than SUMO-OWL, it is likely that it would have performed much better. A similar consideration holds for the use of OpenCyc in our experiments: the full, commercial version of Cyc would have probably given better results. We preferred OpenCyc because it is free and allows everybody to reproduce our experiments.
Acknowledgements We are grateful to A. Pease that provided us with the translation of SUMO into SUMOOWL. This work has been partly supported by the “Iniziativa Software” CINI-FinMeccanica project.
References [1] American National Standard. KIF Knowledge Interchange Format – draft proposed American National Standard (dpANS) NCITS.T2/98-004, 1998. 8
[2] E. Brill, S. Dumais, and M. Banko. An analysis of the askmsr question-answering system. In Conference on Empirical Methods in Natural Language Processing, EMNLP 2002, Proceedings, 2002. [3] N. Casellas, M. Blázquez, A. Kiryakov, P. Casanovas, M. Poblet, and R. Benjamins. OPJK into PROTON: Legal domain ontology integration into an upperlevel ontology. In R. Meersman and et al., editors, WORM 2005, 3rd International Workshop on Regulatory Ontologies, Proceedings, volume 3762 of Lecture Notes in Computer Science, pages 846–855. Springer, 2005. [4] H. H. Do, S. Melnik, and E. Rahm. Comparison of schema matching evaluations. In A. B. Chaudhri, M. Jeckle, E. Rahm, and R. Unland, editors, Web, Web-Services, and Database Systems, NODe 2002 Web and Database-Related Workshops, 2002, Revised Papers, volume 2593 of Lecture Notes in Computer Science, pages 221–237. Springer, 2002. [5] J. Euzenat and P. Shvaiko. Ontology Matching. Springer, 2007. [6] A. Gangemi, N. Guarino, C. Masolo, A. Oltramari, and L. Schneider. Sweetening ontologies with DOLCE. In A. Gómez-Pérez and V. R. Benjamins, editors, Knowledge Engineering and Knowledge Management. Ontologies and the Semantic Web, 13th International Conference, EKAW, Proceedings, volume 2473 of Lecture Notes in Computer Science, pages 166–181. Springer, 2002. [7] P. Grenon, B. Smith, and L. Goldberg. Biodynamic ontology: Applying BFO in the biomedical domain. In D. M. Pisanelli, editor, Ontologies in Medicine, volume 102 of Studies in Health Technology and Informatics, pages 20–38. IOS Press, 2004. [8] H. Herre, B. Heller, P. Burek, R. Hoehndorf, F. Loebe, and H. Michalek. General formal ontology (GFO): A foundational ontology integrating objects and processes. part i: Basic principles. Technical report, Research Group Ontologies in Medicine (Onto-Med), University of Leipzig, 2006. Version 1.0, Onto-Med Report Nr. 8, 01.07.2006. [9] D. Lenat and R. Guha. Building large knowledge-based systems. Addison Wesley, 1990. [10] I. Niles and A. Pease. Towards a standard upper ontology. In C. Welty and B. Smith, editors, FOIS 2001, 2nd International Conference on Formal Ontology in Information Systems, Proceedings, pages 2–9. ACM Press, 2001. [11] A. Pease. The Sigma ontology development environment. In F. Giunchiglia, A. Gomez-Perez, A. Pease, H. Stuckenschmidt, Y. Sure, and S. Willmott, editors, Workshop on Ontology and Distributed Systems, ODS 2003, Proceedings, volume 71 of CEUR-WS, 2003. [12] J. F. Sowa. Knowledge Representation: Logical, Philosophical, and Computational Foundations. Brooks Cole Publishing, 1999.
9
[13] G. Stoilos, G. B. Stamou, and S. D. Kollias. A string metric for ontology alignment. In Y. Gil, E. Motta, V. R. Benjamins, and M. A. Musen, editors, 4th International Semantic Web Conference, ISWC 2005, Proceedings, volume 3729 of Lecture Notes in Computer Science, pages 624–637. Springer, 2005. [14] W3C. OWL Web Ontology Language Overview – W3C Recommendation 10 February 2004 , 2004. [15] Wikipedia. Upper ontology – Wikipedia, the Free Encyclopedia, 2008. [Online; accessed 30-March-2008].
10
URL available from www. ontologyportal.org; the OWL translation performed with Sigma is not available on the Web
SUO-KIF SUMOOWL OpenCyc DOLCE Agent Bibtex Biosphere Ecology Food Geofile HL7_RBAC Ka MPEG7 Restaurant
Resume
Space Subject Top-bio Travel Vacation
Vertebrate
C
O
D
I
4,393
757
2
4,412
26,965
4,854
1
62,469
242
322
4
12
130
75
38
139
15
0
40
2
88
0
0
0
157
35
2
80
65
8
0
57
89
21
4
131
60
24
10
16
96
60
32
0
349
1
0
0
164
27
29
32
167
27
35
46
165
17
4
0
171
0
0
0
65
65
2
1
84
100
112 0
32
10
0
324
19
0
0
0
version
www.cyc.com/2004/06/04/cyc www.loa-cnr.it/ontologies/DLP_ 397.owl 212.119.9.180/Ontologies/0.2/ agent.owl oaei.ontologymatching.org/2004/ Contest/304/onto.rdf sweet.jpl.nasa.gov/ontology/ biosphere.owl wow.sfsu.edu/ontology/rich/ EcologicalConcepts.owl silla.dongguk.ac.kr/jena-owl1/ food www.daml.org/2001/02/geofile/ geofile-ont.daml lsdis.cs.uga.edu/projects/ meteor-s/wsdl-s/ontologies/HL7_ RBAC.owl protege.cim3.net/file/pub/ ontologies/ka/ka.owl dmag.upf.es/ontologies/2003/03/ MPEG7Genres.rdfs guru-games.org/ontologies/ restaurant.owl statistic.gunadarma. ac.id/research/ WorkGroupInformationSystem/ DownLoad/onto_colection/resume. owl 212.119.9.180/Ontologies/0.3/ space.owl www.library.yale.edu/ ontologies/subject.owl www.co-ode.org/ontologies/ basic-bio/top-bio.owl lsdis.cs.uga.edu/projects/ meteor-s/downloads/Lumina/ ontology/travelontology.owl www.guru-games.org/ontologies/ vacation.owl www.co-ode.org/ ontologies/basic-bio/ basic-vertebrate-gross-anatomy. owl *
Table 1: Matched Ontologies; * Vertebrate has been reduced by hand
11
TEST 1 TEST 2
o Ka Biosphere
o’ Bibtex Top-bio
Best Precision OpenCyc, NS Dir, NS; Dir, S
Best Recall SUMO-OWL, M SUMO-OWL, M SUMO-OWL, M; OpenCyc, M OpenCyc, S; OpenCyc, M OpenCyc, M; DOLCE, M SUMO-OWL, M
Best F-measure SUMO-OWL, S SUMO-OWL, M
TEST 3
Space
Geofile
OpenCyc, NS
TEST 4
Restaurant Food
OpenCyc, NS
TEST 5
MPEG7
Subject
OpenCyc, NS
TEST 6
Travel
Vacation
TEST 7
Resume
Agent
Dir, NS; Dir, S SUMO-OWL, NS
OpenCyc, M
OpenCyc, M
TEST 8
Resume
HL7_RBAC
OpenCyc, NS
SUMO-OWL, M
TEST 9
Ecology
Top-bio
Dir, NS; Dir, S
SUMO-OWL, M
TEST 10
Vertebrate
Top-bio
OpenCyc, NS
OpenCyc, S; OpenCyc, M
OpenCyc, M OpenCyc, S; OpenCyc, M SUMO-OWL, NS; OpenCyc, NS SUMO-OWL, M
SUMO-OWL, NS; OpenCyc, NS; OpenCyc, S Dir, NS; Dir, S; SUMO-OWL, M; OpenCyc, M OpenCyc, S
Table 2: Algorithm that gives the best precision, recall and F-measure in each test
Prec. Rec. F-m.
SUMOOWL, NS 0.097 -0.013 -0.011
SUMOOpenCyc, DOLCE, OWL, NS NS S 0.157 -0.145 -0.021 -0.019 -0.027 0.028 -0.018 -0.049 0.000
SUMOOpenCyc, DOLCE, OWL, S S M -0.004 -0.162 -0.089 0.030 -0.031 0.066 0.014 -0.053 0.016
Table 3: Average advantage in using upper ontologies
12
OpenCyc, DOLCE, M M -0.070 0.070 0.024
-0.158 0.020 -0.023
